An OS for companies where AI agents remember last week’s failures

What it does

Agent Swarm is a TypeScript orchestration platform that ingests tasks from Slack, GitHub, GitLab, Linear, Jira, email, or an API. A lead agent breaks down requests and delegates them to worker agents running inside isolated Docker environments. The workers execute the work—shipping pull requests, drafting reports, or answering threads—and write their output and learnings into a persistent memory store so the swarm is meant to accumulate context across sessions.

The interesting bit

The project frames itself as a “Compounding Intelligence Layer,” giving each agent a persistent identity and memory via vector search and per-agent files like SOUL and CLAUDE.md. It also tries to stay harness-agnostic, supporting Claude Code, OpenAI Codex, raw LLMs, and others while preserving task continuity across restarts and provider swaps. That is a harder engineering problem than the usual single-model chat wrapper.

Key highlights

• Lead/worker orchestration inside Docker with priority queues, isolated environments, and pause/resume across deploys.
• Persistent memory and identity system using vector search and agent-specific files, meant to let expertise compound horizontally across the swarm.
• Human-in-the-loop approval gates for irreversible steps, plus workflow patterns like “Litmus Tests” (LLM-as-judge) and “Drain Loops” (chained reviewable PRs).
• Multi-channel task ingestion from Slack, GitHub, email, WhatsApp, Linear, Jira, and more.
• LLM-agnostic harness support with a bounded context preamble that carries task chain history across different providers and restarts.

Caveats

• The README is heavy on vision and light on implementation specifics; most architectural detail lives behind external documentation links.
• The project is backed by Desplega Labs and includes a hosted dashboard and commercial support offers, so self-hosters should expect some vendor ecosystem entanglement.
• A maintainer note warns that the repo “evolves every single day,” which suggests APIs and features are currently in flux.

Verdict

A solid candidate if you are looking to automate complex, cross-tool company workflows and can tolerate early-stage churn. Skip it if you just need a simple LLM wrapper or a one-off cron script.